Deploy Python functions and GPU workloads to Modal serverless cloud infrastructure.
npx skills add https://github.com/tamagusko/linux-cfg --skill modal-tamagusko
Please help me install this Agent Skill. Skill: modal Source: https://github.com/tamagusko/linux-cfg/tree/main/dotfiles/claude/skills/modal Command: npx skills add https://github.com/tamagusko/linux-cfg --skill modal-tamagusko
modal deploy. - Persistent Storage and Secrets: Use Volumes for model weights and datasets, and Secrets for secure credential injection. - Use Case: Serve a 70B LLM as a production API by defining a class with @modal.enter() to load vLLM once on an H100, then expose it via @modal.fastapi_endpoint() with autoscaling containers. ## Quick Start Ask the assistant to write and deploy a Modal app that runs a Python function on a cloud GPU, such as serving an LLM inference endpoint on an H100.High-intent search queries and answers about installing and using this skill.
Install the modal package, authenticate with modal setup, then decorate a function with @app.function(gpu="H100") and call it with .remote(). Modal builds the container and executes the function on the requested GPU in the cloud.
Define a class with @app.cls() that loads the model in @modal.enter(), then expose it with @modal.fastapi_endpoint() or @modal.asgi_app(). Run modal deploy script.py to get a permanent production URL with autoscaling.
L40S is recommended for inference due to its cost-performance balance and 48 GB VRAM. Models above 70B parameters typically need A100-80GB, H100, or multi-GPU configurations like H100:2.
Yes, Modal supports scheduled functions via modal.Cron("0 9 * * *") for cron syntax or modal.Period(hours=6) for fixed intervals. Schedules activate only after running modal deploy, not modal run or modal serve.
Use modal.Volume.from_name("model-weights", create_if_missing=True) and mount it via the volumes parameter. Volumes persist across deployments, support CLI upload with modal volume put, and auto-commit writes in the background.
Modal allows up to 25,000 total inputs, 1,000 concurrent inputs per .map() call, and 200 requests per second on web endpoints. Configure max_containers, min_containers, and @modal.concurrent(max_inputs=N) to control scaling behavior.
wanshuiyinRuns GPU training, inference, and batch workloads on Modal serverless cloud.
george11642Automate deploying and managing Modal workers, logs, and secrets.
DonZzzillaRun local GGUF inference for llama.cpp models with Hugging Face Hub discovery.
BagelHoleDeploy and manage ML models on Kubernetes with KServe and Triton.